S/390: Fix flogr RTX.

The flogr instruction uses a 64 bit register pair target operand.  In
the RTX we model this as a write to a TImode register.  Unfortunately
the RTX's being assigned to the two parts of the target operand were
swapped.  This is no problem if in the end the flogr instruction will
be emitted since the instruction still does what the clzdi expander
expects.  However, a problem arises when the RTX is used to optimize
CLZ for a constant input operand.  Even then it matters only if the
expression couldn't be folded on tree level already.

In the testcase this happened thanks to loop unrolling on RTL level.
The iteration variable is used as an argument to the clz
builtin. Due to the loop unrolling it becomes a constant and after
folding the broken RTX leads to a wrong assumption.

gcc/ChangeLog:

2018-11-20  Andreas Krebbel  <krebbel@linux.ibm.com>

	* config/s390/s390.md ("clztidi2"): Swap the RTX's written to the
	DImode parts of the target operand.

gcc/testsuite/ChangeLog:

2018-11-20  Andreas Krebbel  <krebbel@linux.ibm.com>

	* gcc.target/s390/flogr-1.c: New test.

From-SVN: r266321

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index d9af1e50781a66b00722ecf3f7df0e25a9777ccb..779ed5948787b4221ff36f6658ab39618cd43117 100644 (file)
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,8 @@
+2018-11-20  Andreas Krebbel  <krebbel@linux.ibm.com>
+
+       * config/s390/s390.md ("clztidi2"): Swap the RTX's written to the
+       DImode parts of the target operand.
+
 2018-11-20  Nathan Sidwell  <nathan@acm.org>
 
        PR 87926

diff --git a/gcc/config/s390/s390.md b/gcc/config/s390/s390.md
index 721222d221f4d812fd87319cd76106e1ef923c91..30d113fd9cc94ff8f9b5ab9c03f301908b09e31e 100644 (file)
--- a/gcc/config/s390/s390.md
+++ b/gcc/config/s390/s390.md
@@ -8759,17 +8759,17 @@
   DONE;
 })
 
+; CLZ result is in hard reg op0 - this is the high part of the target operand
+; The source with the left-most one bit cleared is in hard reg op0 + 1 - the low part
 (define_insn "clztidi2"
   [(set (match_operand:TI 0 "register_operand" "=d")
        (ior:TI
-         (ashift:TI
-            (zero_extend:TI
-             (xor:DI (match_operand:DI 1 "register_operand" "d")
-                      (lshiftrt (match_operand:DI 2 "const_int_operand" "")
-                               (subreg:SI (clz:DI (match_dup 1)) 4))))
-
-           (const_int 64))
-          (zero_extend:TI (clz:DI (match_dup 1)))))
+         (ashift:TI (zero_extend:TI (clz:DI (match_operand:DI 1 "register_operand" "d")))
+                    (const_int 64))
+         (zero_extend:TI
+          (xor:DI (match_dup 1)
+                  (lshiftrt (match_operand:DI 2 "const_int_operand" "")
+                            (subreg:SI (clz:DI (match_dup 1)) 4))))))
    (clobber (reg:CC CC_REGNUM))]
   "UINTVAL (operands[2]) == HOST_WIDE_INT_1U << 63
    && TARGET_EXTIMM && TARGET_ZARCH"

diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog
index 4ec625f17689efeca84dc4d2b7919ee30c5e3b2e..d522b2cebdc63211c9fbc470754290187f4489fc 100644 (file)
--- a/gcc/testsuite/ChangeLog
+++ b/gcc/testsuite/ChangeLog
@@ -1,3 +1,7 @@
+2018-11-20  Andreas Krebbel  <krebbel@linux.ibm.com>
+
+       * gcc.target/s390/flogr-1.c: New test.
+
 2018-11-20  Jan Hubicka  <hubicka@ucw.cz>
 
        PR ipa/87706

diff --git a/gcc/testsuite/gcc.target/s390/flogr-1.c b/gcc/testsuite/gcc.target/s390/flogr-1.c
new file mode 100644 (file)
index 0000000..a386900
--- /dev/null
+++ b/gcc/testsuite/gcc.target/s390/flogr-1.c
@@ -0,0 +1,47 @@
+/* { dg-do run } */
+/* { dg-options "-O2 -funroll-loops -march=z9-109" } */
+/* { dg-require-effective-target stdint_types } */
+
+/* Folding of the FLOGR caused a wrong value to be returned by
+   __builtin_clz becuase of a problem in the RTX we emit for FLOGR.
+   The problematic folding can only be triggered with constants inputs
+   introduced on RTL level.  In this case it happens with loop
+   unrolling.  */
+
+#include <stdint.h>
+#include <assert.h>
+
+static inline uint32_t pow2_ceil_u32(uint32_t x) {
+  if (x <= 1) {
+    return x;
+  }
+  int msb_on_index;
+  msb_on_index = (31 ^ __builtin_clz(x - 1));
+  assert(msb_on_index < 31);
+  return 1U << (msb_on_index + 1);
+}
+
+void __attribute__((noinline,noclone))
+die (int a)
+{
+  if (a)
+    __builtin_abort ();
+}
+
+void test_pow2_ceil_u32(void) {
+  unsigned i;
+
+  for (i = 0; i < 18; i++) {
+      uint32_t a_ = (pow2_ceil_u32(((uint32_t)1) << i));
+      if (!(a_ == (((uint32_t)1) << i))) {
+       die(1);
+      }
+  }
+}
+
+int
+main(void) {
+  test_pow2_ceil_u32();
+
+  return 0;
+}